Comment by dilippkumar | Hacker News Reader

Comment by dilippkumar | Hacker News Reader

Parent Full thread

This is a weird question, so I’ll disclaim that I am asking in good faith.

Why is treating GPT-4 as a reasoning machine so controversial? And by that, I don’t mean why it is wrong. I am asking why it upsets so many people.

We ignore a whole bunch of lies and half truths every day. What is it about GPT-4 that doesn’t let it get a free pass like UFOs or EmDrives?

What I am really interested in understanding is the possibility that at least some of these strongly dismissive responses are a bit of a threat response.

The other interesting possibility is that for over a decade, we got used to dismissing AI startups - and now, that habit could persist despite a real breakthrough.

Another possibility is that we were all given an overly simplified model/explanation of machine learning and now, those simplifications are being a real hurdle to comprehension.

It’s extremely fascinating to watch this play out. For my generation, this is the first time we’re seeing a brand new fundamental technology get deployed and alter society.

Speaking for myself, I am equally dismissive of aliens, if that helps.

I think what annoys me about claims like this is two things, and it has to do with in-groups and out-groups.

1. I am concerned about how non-developers perceive us developers when they hear claims like this, since I expect that particular hype balloon to pop in the future.

2. I am distressed to see developers (who should know better) make these types of claims, especially without regard to 1.

If I were more into physics, I'd probably get more worked up about em-drives. As it is, I am enjoying LLMs and consider them an incredibly useful tool, possibly up there with things like GUIs and IDEs in terms of productivity enhancement. But it's still just extra-spicy autocomplete, and I think forgetting that could be dangerous.

Yes you are right that are a kind of threat response but not personal threat as much as general threat. There is a great danger that because of the name AI and the general association that we have with computers as giving correct result, People will (and are already) attribute meaning and "correctness" where there is none. This is nothing like UFOs or EmDrives because those come from fallible humans and this would come from the "infallible AI". What is wrong with describing it for what it is, a stochastic predictive model?

Human brains are known to create "meaning" where there is none and with a perfectly created sentence they will assign meaning (reasoning in this case) where there is none.

One week ago there was a case of a man (in Belgium) who "sacrificed" himself because the AI chatBot convinced him that there was no solution to Enviornment problem. He assigned meaning ("Even the AI could not find a way out") where the chatbot was just parroting what it was reading in the media. Note that in this case both the general media and the AI give the same "result"=no hope, but as long as it was in the general media, he was still hopeful that humans just couldn't find it whereas the AI convinced him.

You may diregard this as an extreme example but we have the general tendency to attribute correctnes to everything "computable". If I can trust the calculator that was created to do math why should I not trust the AI that is created to do Reasoning? People need to be aware that there is no reasoning behind it.

> Why is treating GPT-4 as a reasoning machine so controversial?

At the risk of attracting down votes, you could ask the same question about the visceral responses to crypto currencies you see here.

It's certainly true they have their problems. Bitcoins 7 transactions per second makes it's chain a no-starter for most uses, the numerous multi million dollar losses due to bugs in Ethereum contracts should scare anybody. But the reaction you see here goes well beyond that. After all they continue exist, even performing the occasional useful task, and they are gradually (glacially?) overcoming their initial shortcomings.

The aversion to everything crypto seems to be more of a guttural reaction to the existence of something that uses algorithms to generate and distribute trust. It challenges the human social trust networks we've been using to do the job for centuries. This aversion to calling the obvious deductive capabilities AI's "thinking" or "intelligence" looks to be similar. I guess it challenges what has been a uniquely human ability.

Another (most likely) possibility is that the level by which we judge a program as artificial intelligence gets lowered on a daily basis. What counted as AI didn't count as AI a decade ago.

> and now, that habit could persist despite a real breakthrough

Or the breakthrough has a local maximum, because if LLMs are just statistical token predictors without any real reasoning, it will never achieve AGI level of AI.

Unless, of course, we redefine what an AGI is.

Reply on news.ycombinator.com